fix for #1137/#802, to ensure GSC float and fixed point have the same bit allocation (21b5d7e3) · Commits · IVAS Codec Public Collaboration / IVAS Codec

lib_com/gs_bitallocation.c

+321 −3

Original line number	Diff line number	Diff line
		@@ -53,7 +53,29 @@
		static float Find_bit_frac( const int16_t nb_band, const int16_t remaining_bits );

		static void reajust_bits( float *bits_per_bands, const int16_t st_band, const int16_t end_band, const int16_t sum_bit_in, const int16_t bit_bdgt_in );

		#ifdef NON_BE_FIX_1137_GSC_IVAS_FXFLT_DECODING
		#define Q15_0_33 10922 /* 0.33 */
		#define Q18_0_1 26214 /* 0.1 */
		#define Q18_0_50 131072 /* 0.50 */
		#define Q18_0_75 196608 /* 0.75 */
		#define Q18_0_76 199229 /* 76/100 */
		#define Q18_1_0 262144 /* 1.0 */
		#define Q18_1_2 314573 /* 1.2 */
		#define Q18_112 29360128 /* 112 */
		#define Q18_DSR_NB_PULSE 1179648 /* 4.5 */
		#define Q18_1_5xDSR_NB_PULSE 1769472 /* 1.5x4.5 */
		#define Q18_2_0xDSR_NB_PULSE ( Q18_DSR_NB_PULSE << 1 ) /* 2.0x4.5 */

		#define Q31_0_00125 2684355 /* 0.125/100 */
		#define Q31_0_0125 26843546 /* 0.0125 */
		#define Q31_0_015 32212255 /* 0.0125 */
		#define Q31_0_02 42949673 /* 0.02 */
		#define Q31_0_17 365072220 /* 0.17 */
		#define Q31_0_23 493921239 /* 0.23 */
		static Word16 Find_norm_inv( const Word32 ToDivide, Word16 *e_div );
		static Word16 Find_bit_alloc_IVAS_int( const Word32 core_brate, const Word16 GSC_IVAS_mode, const Word16 Diff_len, const Word16 nb_tot_bands, const Word16 L_frame, Word16 bit, Word16 max_ener_band, float ener_vec, float bits_per_bands );
		static Word16 maximum_fx( const Word16 vec_fx, Word16 lvec_fx, Word16 max_fx );
		#endif

		/-------------------------------------------------------------------
		* bands_and_bit_alloc()
		@@ -92,9 +114,12 @@ void bands_and_bit_alloc(
		int16_t pos, band;
		float SWB_bit_budget;
		float bits_per_bands[MBANDS_GN_BITALLOC16k];
		#ifndef NON_BE_FIX_1137_GSC_IVAS_FXFLT_DECODING
		float fzero_val, mp, mb, nb_bands_adj, bit_adj;
		int16_t nb_pulse_per_band[MBANDS_GN_BITALLOC16k];

		#else
		float fzero_val;
		#endif
		/* initializations */
		nb_tot_bands = 16;
		set_f( bits_per_bands, 0.0f, MBANDS_GN_BITALLOC16k );
		@@ -193,6 +218,7 @@ void bands_and_bit_alloc(
		{
		if ( GSC_IVAS_mode > 0 )
		{
		#ifndef NON_BE_FIX_1137_GSC_IVAS_FXFLT_DECODING
		SWB_bit_budget = *bit;
		st_band = 5;

		@@ -371,6 +397,10 @@ void bands_and_bit_alloc(
		bits_per_bands[i] += sum_bit;
		}
		}
		#else
		nb_tot_bands = Find_bit_alloc_IVAS_int( core_brate, GSC_IVAS_mode, Diff_len, nb_tot_bands, L_frame, bit, max_ener_band, ener_vec, bits_per_bands );
		nb_bands = nb_tot_bands;
		#endif
		}
		else if ( GSC_noisy_speech )
		{
		@@ -824,7 +854,6 @@ void bands_and_bit_alloc(
		*nb_subbands = 0;
		*pvq_len = 0;
		}

		return;
		}

		@@ -933,3 +962,292 @@ static float Find_bit_frac(

		return ( var_out );
		}
		#ifdef NON_BE_FIX_1137_GSC_IVAS_FXFLT_DECODING

		static Word16 Find_bit_alloc_IVAS_int( /o: Number of band to encode /
		const Word32 core_brate, /* i : core bit rate */
		const Word16 GSC_IVAS_mode, /* i : GSC IVAS mode */
		const Word16 Diff_len, /* i : Length of the difference signal (before pure spectral)*/
		const Word16 nb_tot_bands_in, /* i : total number of band */
		const Word16 L_frame, /* i : frame length */
		Word16 bit, / i/o: Number of bit allowed for frequency quantization */
		Word16 max_ener_band, / i/o: Energy based sorted order */
		float ener_vec_io, / i/o: Energy per band order */
		float bits_per_bands_o / o : Number of bit allowed per allowed sub-band Q3 */
		)
		{
		Word32 mp, mb, nb_bands_adj, bit_adj;
		Word16 nb_pulse_per_band[MBANDS_GN_BITALLOC16k];
		Word32 SWB_bit_budget; // Q0 -> Q18
		Word16 i, j, nb_bands_max, st_band, nb_tot_bands_loc, etmp;
		Word32 sum_bit /Q18/, bit_fracf /Q18/;
		Word16 d_tmp, e_div, tmp16, ener_vec[MBANDS_GN_BITALLOC16k];
		Word32 Ltmp, etmp_32fx, bits_per_bands[MBANDS_GN_BITALLOC16k];


		SWB_bit_budget = *bit; // Q0
		st_band = 5;
		nb_bands_max = nb_tot_bands_in;


		for ( i = 0; i < MBANDS_GN; i++ )
		{
		ener_vec[i] = (short) ( ener_vec_io[i] ); /* Q12 -> Q12 */
		}
		if ( L_frame == L_FRAME16k )
		{
		for ( i = MBANDS_GN; i < MBANDS_GN_BITALLOC16k; i++ )
		{
		ener_vec[i] = (short) ( ener_vec_io[i] * 4096.0 + 0.5f ); /* Q0 -> Q12 */
		}
		}

		set_l( bits_per_bands, 0, MBANDS_GN_BITALLOC16k );

		/* Decide the pourcentage of bits allocated to LF (between 50-75%) depending of the temporal contribution in GSC */
		/* bit_fracf = ( -0.125f * Diff_len + 76.0f ) / 100; */
		bit_fracf = L_add( Mpy_32_32( -Q31_0_00125, L_shl( Diff_len, 18 ) ), Q18_0_76 ); /* Q18 */

		/* bit_fracf = check_bounds(bit_fracf, 0.50f, 0.75f); */
		bit_fracf = min( max( bit_fracf, Q18_0_50 ), Q18_0_75 );

		/* Adjusment of the bitrate between LF and HF base on the content type */
		/* 1 = new GSC bit alloc
		2 = GSC bit alloc for tc frame
		3 = more music like (should not happen often given music is coded with dft) */

		if ( GSC_IVAS_mode <= 3 )
		{
		nb_bands_max -= 6;
		}
		if ( GSC_IVAS_mode == 2 )
		{
		/* bit_fracf += 0.1f; */
		bit_fracf += Q18_0_1; /* Q18*/
		nb_bands_max -= 1;
		}
		if ( GSC_IVAS_mode == 3 )
		{
		/* bit_fracf -= 0.1f; */
		bit_fracf -= Q18_0_1; /* Q18*/
		nb_bands_max += 3;
		}

		/* First find how much we want to share between LF and HF, at low bitrate, a miminum of bits is needed in LF by limitating the number of bands*/
		/* Adjust the number of band based on the content type and bitrate */

		/* nb_bands_adj = 1.0f; */
		nb_bands_adj = Q18_1_0;
		if ( GSC_IVAS_mode == 1 && core_brate < GSC_L_RATE_STG )
		{
		/* nb_bands_adj = 0.0125f * SWB_bit_budget - 0.75f;*/
		nb_bands_adj = L_sub( Mpy_32_32( Q31_0_0125, L_shl( SWB_bit_budget, 18 ) ), Q18_0_75 ); // Q18
		}
		else if ( GSC_IVAS_mode != 2 && core_brate > GSC_H_RATE_STG )
		{
		/nb_bands_adj = 0.02f SWB_bit_budget - 1.2f;*/
		nb_bands_adj = L_sub( Mpy_32_32( Q31_0_02, L_shl( SWB_bit_budget, 18 ) ), Q18_1_2 ); // Q18
		}
		/nb_bands_max = (int16_t)(nb_bands_max nb_bands_adj + 0.5f);*/
		nb_bands_max = (Word16) L_shr_r( nb_bands_max * nb_bands_adj, 18 ); /* Q0 */ ///################################
		nb_bands_max = check_bounds_s( nb_bands_max, 5, nb_tot_bands_in );

		/* bit_fracf = SWB_bit_budget;/
		/* At this point bit_fracf has a value below 1.0 */
		bit_fracf = Mpy_32_16_1( L_shl( bit_fracf, 10 ), extract_l( L_shl( SWB_bit_budget, 5 ) ) ); /* (Q(18+10)Q(0+5) + 1 - 16 = Q18 /

		/* Estimation of the number of bit used in HF */
		/* with only the first weighting The number of bits in max_ener_band[st_band-1] = 17% of bit_fracf */
		/* mb = .17f * bit_fracf;*/
		mb = Mpy_32_32( Q31_0_17, bit_fracf ); /* Q18 */

		/* mp = 2 * DSR_NB_PULSE;*/
		mp = Q18_2_0xDSR_NB_PULSE;
		if ( core_brate < GSC_L_RATE_STG && GSC_IVAS_mode == 3 )
		{
		/* mp = 1.5f * DSR_NB_PULSE;*/
		mp = Q18_1_5xDSR_NB_PULSE;
		}
		else if ( core_brate < GSC_L_RATE_STG )
		{
		/* mp = DSR_NB_PULSE;*/
		mp = Q18_DSR_NB_PULSE;
		}

		/* We want max_ener_band[st_band] <= max_ener_band[st_band-1] and max_ener_band[nb_bands_max-1] <= max_ener_band[st_band]*/
		/* We will estimate the number of bits to allocate of HF and put the remaining bits, if any, back on LF */
		/* compute the total possible number of band to be coded */

		/* nb_tot_bands = (int16_t)((SWB_bit_budget - bit_fracf) / (mp + (mb - mp) / 2.0f)); */
		d_tmp = Find_norm_inv( L_add( mp, mb ), &e_div );
		Ltmp = Mpy_32_16_1( L_sub( L_shl( SWB_bit_budget, 18 ), bit_fracf ), d_tmp ); /* Perform mult by 1/den */
		nb_tot_bands_loc = extract_h( L_shl( Ltmp, sub( 1, e_div ) ) ); /* adjust exponent: 1 is to take into account the / 2.0f, and e_div for the num and den of the division*/

		mp = min( mp, mb );
		tmp16 = sub( add( nb_tot_bands_loc, st_band ), nb_bands_max );
		if ( tmp16 > 0 )
		{
		/* bit_adj = ( ( mb + mp ) / 2 ) * ( nb_tot_bands_loc + st_band - nb_bands_max ); */
		bit_adj = Mpy_32_16_1( L_shl( L_add( mb, mp ), 5 ), shl( tmp16, 10 - 1 ) ); /* Q18+5 * Q0+10 + 1 -1 - 16 = Q18 (-1 is to cover for the /2 in the equation) */
		bit_adj = L_max( 0, bit_adj );
		nb_tot_bands_loc = nb_bands_max - st_band;
		bit_fracf += bit_adj; /* Q18 */
		}
		nb_tot_bands_loc += st_band;

		/* Allocate bits to LF */
		/* etmp = 0.23f; */
		etmp_32fx = Q31_0_23;
		for ( j = 0; j < st_band; j++ )
		{
		i = j;
		max_ener_band[j] = i;
		ener_vec[i] = MIN16B;
		/* bits_per_bands[j] = etmp * bit_fracf; */
		bits_per_bands[j] = Mpy_32_32( bit_fracf, etmp_32fx ); /* 18 + 31 + 1 - 32 = Q18 */
		/* etmp -= 0.015f; */
		etmp_32fx -= Q31_0_015; /* Q18 */
		}

		/* SWB_bit_budget -= bit_fracf; */
		SWB_bit_budget = L_sub( L_shl( SWB_bit_budget, 18 ), bit_fracf ); /* Q0->Q18 */

		/* Find low energy band in HF */
		set_s( nb_pulse_per_band, 2, MBANDS_GN_BITALLOC16k );
		for ( i = st_band + 2; i < nb_tot_bands_loc - 1; i++ )
		{
		if ( ener_vec[i] < ener_vec[i - 1] && ener_vec[i] < ener_vec[i + 1] ) /* i +1 and i -1 can be considered as 2 ptrs */
		{
		nb_pulse_per_band[i] = 1;
		}
		}
		for ( j = st_band; j < nb_tot_bands_loc; j++ )
		{
		if ( j > 6 )
		{
		i = maximum_fx( ener_vec, nb_tot_bands_loc, &etmp );
		}
		else
		{
		i = j;
		}
		max_ener_band[j] = i;
		ener_vec[i] = MIN16B;
		}
		/* Recompute the final bit distribution for HF */
		IF( nb_tot_bands_loc > st_band )
		{
		/* This is not bit exact because of the precision lost */
		/* mb = ( SWB_bit_budget * 2 / ( nb_tot_bands_loc - st_band ) ) - mp; */
		d_tmp = Find_norm_inv( L_deposit_h( sub( nb_tot_bands_loc, st_band ) ), &e_div );
		mb = L_sub( L_shr( Mpy_32_16_1( L_shl( SWB_bit_budget, 1 ), d_tmp ), e_div ), mp ); /* Q18 */
		/* bit_fracf = ( mb - mp ) / ( nb_tot_bands_loc - st_band ); */
		bit_fracf = L_shr( Mpy_32_16_1( L_sub( mb, mp ), d_tmp ), e_div ); /* Q18 */

		mb -= bit_fracf;
		/* Do the distribution */
		for ( j = st_band; j < nb_tot_bands_loc; j++ )
		{
		bits_per_bands[max_ener_band[j]] = Q18_DSR_NB_PULSE;
		if ( nb_pulse_per_band[max_ener_band[j]] > 1 )
		{
		bits_per_bands[max_ener_band[j]] = mb;
		}
		mb -= bit_fracf;
		SWB_bit_budget -= bits_per_bands[max_ener_band[j]]; // Q18
		}
		}
		/* Series of verification in case bit allocated != the budget */
		if ( SWB_bit_budget > 0 )
		{
		i = st_band - 1;
		while ( SWB_bit_budget > 0 )
		{
		/* bits_per_bands[i]++; */
		bits_per_bands[i] += Q18_1_0;
		/* SWB_bit_budget--; */
		SWB_bit_budget -= Q18_1_0;
		i--;
		if ( i == -1 )
		{
		i = st_band - 1;
		}
		}
		}

		/nb_bands = nb_tot_bands_loc;/

		sum_bit = 0;
		j = 0;
		for ( i = 0; i < nb_tot_bands_loc; i++ )
		{
		/* if (bits_per_bands[i] > 112) */
		if ( bits_per_bands[i] > Q18_112 )
		{
		/* sum_bit += bits_per_bands[i] - 112; */
		sum_bit = L_add( sum_bit, L_sub( bits_per_bands[i], Q18_112 ) );
		/* bits_per_bands[i] = 112; */
		bits_per_bands[i] = Q18_112;
		j = add( j, add( i, 1 ) );
		}

		/* safety check for overage bit reallocation */
		/* else if (bits_per_bands[i] + sum_bit / 3 > 112) */
		else if ( L_add( bits_per_bands[i], Mpy_32_16_1( sum_bit, Q15_0_33 ) ) > Q18_112 )
		{
		j = add( j, add( i, 1 ) );
		}
		}

		if ( sum_bit != 0 )
		{
		/* sum_bit /= (nb_bands - j); */
		d_tmp = Find_norm_inv( L_deposit_h( sub( nb_tot_bands_loc, j ) ), &e_div );
		sum_bit = L_shr( Mpy_32_16_1( sum_bit, d_tmp ), e_div ); /* Q18 */
		for ( i = j; i < nb_tot_bands_loc; i++ )
		{
		bits_per_bands[i] = L_add( bits_per_bands[i], sum_bit );
		}
		}
		for ( i = 0; i < MBANDS_GN_BITALLOC16k; i++ )
		{
		bits_per_bands_o[i] = (float) bits_per_bands[i] / 262144.0f; /* Q18 -> float */
		}
		return nb_tot_bands_loc;
		}

		static Word16 Find_norm_inv( const Word32 ToDivide, Word16 e_div ) / Find normalized 1 / ToDivide */
		{
		Word16 d_tmp, e_tmp;
		e_tmp = norm_l( ToDivide );
		d_tmp = round_fx( L_shl( ToDivide, e_tmp ) );
		d_tmp = div_s( 16384, d_tmp ); /* 1.0 in Q14, dividend is normalize so >= 16384 as required for the division */
		*e_div = sub( 14, e_tmp );
		move16();
		return d_tmp;
		}
		static Word16 maximum_fx( /* o : index of the maximum value in the input vector */
		const Word16 vec_fx, / i : input vector */
		const Word16 lvec_fx, /* i : length of input vector */
		Word16 max_fx / o : maximum value in the input vector */
		)
		{
		Word16 j, ind;
		Word16 tmp;
		ind = 0;

		tmp = vec_fx[0];

		for ( j = 1; j < lvec_fx; j++ )
		{
		if ( vec_fx[j] > tmp )
		{
		ind = j;
		}
		tmp = s_max( tmp, vec_fx[j] );
		}
		*max_fx = tmp;

		return ind;
		}
		#endif

lib_com/options.h

+2 −1

Original line number	Diff line number	Diff line
		@@ -179,6 +179,7 @@
		#define NONBE_FIX_1128_OSBA_EXT_OUTPUT_ORDER /* FhG: issue 1128: set output ambisonics order to input order for EXT output */
		#define NONBE_FIX_1052_SBA_EXT_FIX /* VA: SBA external output support fix - do not overwrite "output_config" parameter */
		#define NONBE_FIX_1130_DIV_ZERO_LEV_DUR /* VA: issue 1130: avoid div by zero in L-D by thresholding R[0] to a min value of 100.0 */
		#define NON_BE_FIX_1137_GSC_IVAS_FXFLT_DECODING /* VA: Add fix point bit allocation for special GSC mode such that float and fixed point have the same final bit allocation */

		/* ##################### End NON-BE switches ########################### */